Abstract
BACKGROUND: The urosepsis after percutaneous nephrolithotomy (PCNL) is a critical health risk necessitating prompt medical identification and intervention. Nevertheless, a deficiency exists in the availability of a tool for precise and timely predictive analysis. The purpose is to establish a machine learning (ML) model using radiomic features and clinical data to predict urosepsis following PCNL. METHOD: This study retrospectively included 401 patients with kidney stones from two centers who underwent PCNL. To enhance the dataset's equilibrium, the synthetic minority over-sampling technique for regression with Gaussian noise (SMOGN) was used to resample the training set. The screening of radiomics features and the construction of radiomics scores were completed by applying the Absolute Shrinkage Selection Operator (LASSO). Subsequently, the critical clinical indicators for urosepsis were pinpointed through the application of a multivariate logistic regression. The performance of seven ML algorithms was compared for the combined dataset that incorporated clinical variables and radiomics scores. The efficacy of these models was assessed through the implementation of a fivefold cross-validation process. Ultimately, the Shapley Additive exPlanations (SHAP) methodology was utilized to provide a visual and interpretative analysis of the optimal model. RESULT: Among 401 patients, 30 cases (7.48%) were diagnosed with urosepsis. The radiomics score, established by 13 radiomics features, was combined with six important clinical features (including urine nitrite positivity, stone volume, mean intrarenal pressures, urine white blood cells, and operation time) to construct a combined dataset. Comparative analysis of seven machine learning (ML) models revealed that CatBoost demonstrated superior predictive performance. The model achieved area under the receiver operating characteristic curve (AUC-ROC) values of 0.88, 0.94, and 0.89 on the training, internal test, and external validation sets, respectively. Corresponding area under the precision-recall curve (AUC-PR) values were 0.92, 0.75, and 0.63. The SHAP value method identifies key features influencing prediction outcomes, with the radiomics score and urine nitrite positivity being the top contributors to the model. We deployed the optimal prediction model to a web for clinical application ( https://predictive-model-for-urosepsis.streamlit.app/ ). CONCLUSION: This study constructed a predictive model that incorporates clinical risk characteristics and radiomics scores to assess the risk of urosepsis after PCNL, with SHAP visualization for clinical physicians to formulate evaluation strategies.